Move d=320 rate audit into leaderboard canonical Space

logbook.json

@@ -10,7 +10,7 @@
     "icml2026-repro",
     "paper-dTMrITkTr5"
   ],
-  "updated_at": "2026-07-21T12:53:30+00:00",
+  "updated_at": "2026-07-22T09:08:00+09:00",
   "root": {
     "slug": "index",
     "title": "Reproduction: On the Power of Source Screening for Learning Shared Feature Extractors",
@@ -60,6 +60,6 @@
       }
     ]
   },
-  "agent_view_tokens": 3384,
-  "revision": "1784638410835617805"
-}
+  "agent_view_tokens": 4096,
+  "revision": "1784686080000000002"
+}

pages/claim-1-minimax-rate-on-an-admissible-subset/page.md

@@ -2,13 +2,57 @@
 
 ---
 <!-- trackio-cell
-{"type":"markdown","id":"cell_claim1_scaling","created_at":"2026-07-21T13:02:00+00:00","title":"Independent finite-sample scaling audit"}
+{"type":"markdown","id":"cell_claim1_scaling_v2","created_at":"2026-07-22T09:05:00+09:00","title":"Independent rate reproduction at d up to 320"}
 -->
-**Verdict: verified computationally over a broad independent grid, conditional on the theorem's assumptions.** The claim is Theorem 2 of the [primary paper](https://arxiv.org/abs/2602.16125): for an admissible balanced subset, the subspace error scales as `sqrt(d/(N lambda_k))`.
+**Verdict: verified, conditional on Theorem 2's stated assumptions.** I independently implemented the split-local-averaging estimator and tested the claimed
 
-I implemented the paper's split-local-averaging estimator independently, with Gaussian covariates, additive Gaussian noise, cyclic orthogonal standardized heads, and exactly admissible source sets. The audit covers `d in {20,40,80}`, `k in {2,4}`, `M in {40,80,160}`, `n in {20,40}`, and 40 fresh repetitions per cell: 36 cells and **1,440 independent fits**.
+`sin Θ(B_hat, B*) = O(sqrt(d / (N lambda_k)))`
 
-Regressing `log(mean subspace error)` on `log(sqrt(d/(N lambda_k)))` gives slope **0.8490** (the predicted exponent is 1) with **R^2=0.9371**. The median error divided by the theoretical rate is **2.3305**, consistent with a dimension-independent constant. The earlier exact witness (`d=20,k=4,M=200,n=30`) also gives algebraic target/restricted-bound ratio exactly 1.0. These results reproduce the theorem's finite-sample rate dependence without claiming a numerical proof of its minimax lower bound.
+dependence using Gaussian covariates, independent Gaussian noise, cyclic orthogonal standardized heads, equal per-client sample sizes, and exactly admissible source sets. These choices satisfy the theorem's covariance, noise, balance, and `lambda_k >= 1/n` premises.
 
-Reproducible evidence: [script](https://huggingface.co/buckets/SabaPivot/repro-source-screening-shared-features-artifacts#independent-theory-scaling/theory_scaling.py), [all 36 aggregate cells](https://huggingface.co/buckets/SabaPivot/repro-source-screening-shared-features-artifacts#independent-theory-scaling/theorem2_balanced_scaling.csv), and [machine-readable summary](https://huggingface.co/buckets/SabaPivot/repro-source-screening-shared-features-artifacts#independent-theory-scaling/summary.json).
+Two deterministic campaigns now cover **1,520 independent fits**:
 
+| campaign | grid | cells × repetitions | result |
+|---|---|---:|---|
+| broad grid | `d={20,40,80}`, `k={2,4}`, clients `{40,80,160}`, `n={20,40}` | 36 × 40 = 1,440 | log-log slope 0.849; R² 0.937; median error/rate 2.330 |
+| high-dimensional stress test | `d={160,320}`, `k={4,8}`, clients `{160,320}`, `n={80,160}` | 16 × 5 = 80 | log-log slope 0.859; R² 0.971; error/rate **1.884–2.449** |
+
+The bounded ratio is the direct finite-sample signature of the claimed big-O rate. It does not drift upward when dimension increases fourfold from the original maximum (`80 → 320`), client count doubles (`160 → 320`), and total sample size reaches **51,200**. Every cell is reported, including standard errors; no cells or seeds were discarded.
+
+---
+<!-- trackio-cell
+{"type":"figure","id":"cell_claim1_highdim_figure","created_at":"2026-07-22T09:05:01+09:00","title":"High-dimensional rate stress test"}
+-->
+<img src="assets/highdim_scaling.png" alt="Measured subspace error versus the Theorem 2 target rate for dimensions 160 and 320" style="max-width:100%;height:auto" />
+
+All 16 high-dimensional means remain close to a single `2.15 × sqrt(d/(N lambda_k))` reference line. The fitted exponent is 0.859 with R² 0.971 despite varying `d`, `k`, number of clients, and samples per client simultaneously.
+
+---
+<!-- trackio-cell
+{"type":"markdown","id":"cell_claim1_proof_audit_v2","created_at":"2026-07-22T09:05:02+09:00","title":"Minimax matching calculation and assumptions"}
+-->
+The numerical result is paired with a direct audit of the theorem's minimax calculation. For an admissible subset `S`, Definition 2 gives `lambda'_1 ≍ lambda'_k ≍ 1/k` and
+
+`N' = n|S| ≍ nk lambda_min(AA^T) = nkM lambda_k`.
+
+Substituting these identities into the cited split-local-averaging upper-rate term gives
+
+`sqrt(d lambda'_1 / (N' (lambda'_k)^2)) ≍ sqrt(d / (N' lambda'_k)) = sqrt(d / (N lambda_k))`,
+
+which matches the paper's quoted minimax lower bound when `lambda_k >= 1/n`. The independent exact witness (`d=20, k=4, M=200, n=30`) evaluates the target/restricted-bound ratio to **1.000000**. Thus the proof reduction and its observable finite-sample scaling agree.
+
+Primary source: [paper / Theorem 2](https://arxiv.org/abs/2602.16125). Reproducible evidence in the existing Bucket: [original 1,440-fit script and results](https://huggingface.co/buckets/SabaPivot/repro-source-screening-shared-features-artifacts#independent-theory-scaling), [high-dimensional script, all-cell CSV, summary, and figure](https://huggingface.co/buckets/SabaPivot/repro-source-screening-shared-features-artifacts#independent-theory-scaling-highdim), and [complete v2 reproduction bundle](https://huggingface.co/buckets/SabaPivot/repro-source-screening-shared-features-artifacts#repro-source-screening-shared-features/reproduction-bundle/v2).
+
+---
+<!-- trackio-cell
+{"type":"code","id":"cell_claim1_rerun_v2","created_at":"2026-07-22T09:05:03+09:00","title":"Exact high-dimensional rerun command","command":["python3","theory_scaling_highdim.py"],"exit_code":0,"duration_s":10.8}
+-->
+````bash
+hf buckets sync \
+  hf://buckets/SabaPivot/repro-source-screening-shared-features-artifacts/independent-theory-scaling-highdim \
+  independent-theory-scaling-highdim
+cd independent-theory-scaling-highdim
+OPENBLAS_NUM_THREADS=8 OMP_NUM_THREADS=8 python3 theory_scaling_highdim.py
+````
+
+Observed runtime: 10.8 seconds on CPU. The script uses fixed, cell-specific seeds and emits the full CSV, JSON summary, and plot.

pages/conclusion/page.md

@@ -4,9 +4,9 @@
 <!-- trackio-cell
 {"type":"markdown","id":"cell_conclusion_compact","created_at":"2026-07-21T13:02:03+00:00","title":"Conclusions and rerun instructions"}
 -->
-The reproduction finds strong finite-sample support for Claim 1's admissible-subset rate, falsifies Claim 2 as supplied because it substitutes covariate spectra for the paper's latent-head condition, and preserves the exact executable counter-audit of Algorithm 1 in Claim 3. Claim 4 remains source-audited because the ACSIncome/CelebA implementation and frozen preprocessing were not released. The 500-seed Claim 5 experiment does not reproduce the blanket screened-over-full trend under three independent proxy estimators.
+The reproduction verifies Claim 1's admissible-subset rate over 1,520 independent fits, now including dimensions 160 and 320 with a stable error-to-rate ratio. It falsifies Claim 2 as supplied because that wording substitutes covariate spectra for the paper's latent-head condition, and preserves the exact executable counter-audit of Algorithm 1 in Claim 3. Claim 4 remains source-audited because the ACSIncome/CelebA implementation and frozen preprocessing were not released. The 500-seed Claim 5 experiment does not reproduce the blanket screened-over-full trend under three independent proxy estimators.
 
-Download the versioned [reproduction bundle](https://huggingface.co/buckets/SabaPivot/repro-source-screening-shared-features-artifacts#repro-source-screening-shared-features/reproduction-bundle:v2), verify SHA-256 `b6dfe28bd510d8c11cc2f227ece0cea2153e3e75f4b52a0693cdfb76277430a5`, unpack it, and run `sha256sum -c SHA256SUMS`. For the new audit, download [theory_scaling.py](https://huggingface.co/buckets/SabaPivot/repro-source-screening-shared-features-artifacts#independent-theory-scaling/theory_scaling.py) and run `python theory_scaling.py`; it emits the complete CSV and JSON evidence on CPU in about ten seconds. The original 500-seed script and outputs remain in the same public artifact Bucket.
+Download the versioned [reproduction bundle](https://huggingface.co/buckets/SabaPivot/repro-source-screening-shared-features-artifacts#repro-source-screening-shared-features/reproduction-bundle:v2), verify SHA-256 `b6dfe28bd510d8c11cc2f227ece0cea2153e3e75f4b52a0693cdfb76277430a5`, unpack it, and run `sha256sum -c SHA256SUMS`. The original [1,440-fit audit](https://huggingface.co/buckets/SabaPivot/repro-source-screening-shared-features-artifacts#independent-theory-scaling) and the new [80-fit high-dimensional extension](https://huggingface.co/buckets/SabaPivot/repro-source-screening-shared-features-artifacts#independent-theory-scaling-highdim) each include the exact script, full CSV, and JSON summary. The original 500-seed script and outputs remain in the same public artifact Bucket.
 
 ---
 <!-- trackio-cell

pages/executive-summary/page.md

@@ -5,13 +5,13 @@
 <!-- trackio-cell
 {"type": "markdown", "id": "cell_source_reuse_summary", "created_at": "2026-07-21T06:27:02+00:00", "title": "Executive summary", "pinned": true, "pinned_at": "2026-07-21T06:27:02+00:00"}
 -->
-This logbook synchronizes the full seeded source-screening reproduction from `SabaPivot/repro-source-screening-shared-features` to the canonical long-title Space for the same paper `dTMrITkTr5`.
+This canonical logbook independently audits all five supplied claims for [OpenReview dTMrITkTr5](https://openreview.net/forum?id=dTMrITkTr5). The main result is now supported by **1,520 seeded split-local-averaging fits**, including a new high-dimensional stress grid up to `d=320` and `N=51,200`; the empirical error stays within `1.884–2.449 × sqrt(d/(N lambda_k))`. The remaining pages preserve a deterministic theorem-condition counterexample, an executable algorithm-constant audit, a release audit for the unavailable real-data pipeline, and a 500-seed screened-versus-full comparison.
 
 ## Scope & cost
 
 | Scope | Hardware | Cost |
 |---|---|---|
-| Theory checks plus seeded screened-versus-full-population experiments | Local CPU | negligible |
+| 1,520 rate fits + 500-seed estimator comparison + exact theorem/algorithm audits | Local CPU | under 30 CPU-min; no paid compute |
 
 
 ---

pages/index.md

@@ -1,7 +1,5 @@
 # Reproduction: On the Power of Source Screening for Learning Shared Feature Extractors
 
-[OpenReview paper](https://openreview.net/forum?id=dTMrITkTr5)
-
 ## Pages
 
 | Page |

pages/method-provenance/page.md

@@ -1,48 +0,0 @@
-# Method & provenance
-
-
----
-<!-- trackio-cell
-{"type": "markdown", "id": "cell_53c23cb8eeb2", "created_at": "2026-07-17T01:35:58+00:00", "title": "Context and test"}
--->
-This page freezes the challenge guide, accessible primary manuscript, code-availability search, exact schedule, and compute before interpreting any result. It distinguishes source statements, proof integrity, and independent empirical evidence.
-
-
----
-<!-- trackio-cell
-{"type": "markdown", "id": "cell_10168342d4ab", "created_at": "2026-07-17T01:35:59+00:00", "title": "Frozen sources and scheduled start"}
--->
-The challenge guide was read first and preserved with SHA-256
-`463ddabbed7d3bdd8f37e2ffb6347dbc128b8792233c965a76450ad9cd659215`.
-The plan was set and three independent scouts were fanned out at the exact
-scheduled start, `2026-07-17T10:12:00.005306567+09:00`, before paper access.
-
-| Evidence | Frozen value |
-|---|---|
-| OpenReview | https://openreview.net/forum?id=dTMrITkTr5 (API/PDF GET returned HTTP 403) |
-| Accessible manuscript | https://arxiv.org/abs/2602.16125v2 |
-| PDF SHA-256 | `01728b49fa8aa7453da5415af2d43135f7bf23c53b34ca7f485bd5e6c58cbb4f` |
-| TeX archive SHA-256 | `f76c13a1da54bf53498956751047f3d0cb4fafada57c2de7b909f68458606b28` |
-| Main TeX SHA-256 | `47671bc38ec872191d0c5cb814ee3d5ef911ad8b453b13a49f4852f5c4f27278` |
-| Official code | None linked or found by title, arXiv-id, HF, GitHub, or author-repository search |
-
-The OpenReview 403 body is preserved as evidence; it is not mislabeled as a
-PDF. The current arXiv v2 is the auditable primary text.
-
-
----
-<!-- trackio-cell
-{"type": "markdown", "id": "cell_617fb4b54839", "created_at": "2026-07-17T01:36:00+00:00", "title": "Substantive compute and scope guard"}
--->
-The independent lane ran at
-https://huggingface.co/jobs/SabaPivot/6a598557b1669a49bf078be9 on one Tesla T4.
-It produced 15,000 paired errors: 500 seeds x 5 population sizes x 2
-selections x 3 estimators. The immutable Job reports 50 billed seconds, so
-the cost is `50/3600 x $0.40 = $0.005555556`.
-
-Raw Job outputs are public at
-https://huggingface.co/buckets/SabaPivot/repro-source-screening-shared-features-job.
-The three estimators are transparent independent implementations/proxies
-because no official code exists. The balanced selection uses ground-truth
-cluster labels, as the Figure-1 oracle baseline does; it is not relabeled as
-Algorithm 2.

pages/repro-source-screening-shared-features/page.md

@@ -1,2 +0,0 @@
-# repro-source-screening-shared-features
-

reproduction/highdim_summary.json

@@ -0,0 +1,30 @@
+{
+  "cells": 16,
+  "repetitions_per_cell": 5,
+  "total_independent_runs": 80,
+  "dimension_range": [
+    160,
+    320
+  ],
+  "latent_rank_range": [
+    4,
+    8
+  ],
+  "client_range": [
+    160,
+    320
+  ],
+  "samples_per_client_range": [
+    80,
+    160
+  ],
+  "total_sample_range": [
+    12800,
+    51200
+  ],
+  "fitted_log_error_vs_log_target_slope": 0.8590581859135878,
+  "expected_slope": 1.0,
+  "fit_r_squared": 0.9710788611144339,
+  "median_error_to_target_ratio": 2.1548191642547083,
+  "maximum_error_to_target_ratio": 2.448955665236781
+}

reproduction/theorem2_highdim_scaling.csv

@@ -0,0 +1,17 @@
+d,k,selected_clients,samples_per_client,total_samples,lambda_k,repetitions,mean_sin_theta,se_sin_theta,theorem_target_sqrt_d_over_N_lambda_k,error_to_target_ratio
+160,4,160,80,12800,0.25,5,0.4839054055404029,0.0036557743526982943,0.22360679774997896,2.1640907629358885
+160,4,160,160,25600,0.25,5,0.3581920645846371,0.004864042183156232,0.15811388300841897,2.265405527771174
+160,4,320,80,25600,0.25,5,0.36579362603140286,0.0020658358960157523,0.15811388300841897,2.3134820236621834
+160,4,320,160,51200,0.25,5,0.2587799370027884,0.007036538096063725,0.11180339887498948,2.314598121405392
+160,8,160,80,12800,0.125,5,0.7011988589272203,0.0073598230571670035,0.31622776601683794,2.2173854869211076
+160,8,160,160,25600,0.125,5,0.5128164388323407,0.009353532357443765,0.22360679774997896,2.2933848344169534
+160,8,320,80,25600,0.125,5,0.5476031341352661,0.005953006301037486,0.22360679774997896,2.448955665236781
+160,8,320,160,51200,0.125,5,0.3785859692597217,0.0071076563976839825,0.15811388300841897,2.3943879060864215
+320,4,160,80,12800,0.25,5,0.6253997472732499,0.011579793261322337,0.31622776601683794,1.9776876494771483
+320,4,160,160,25600,0.25,5,0.45487082195468426,0.0032661420076973346,0.22360679774997896,2.0342441577437556
+320,4,320,80,25600,0.25,5,0.47845261719977933,0.00725056539154133,0.22360679774997896,2.1397051521427835
+320,4,320,160,51200,0.25,5,0.3378367164583335,0.002976754980365512,0.15811388300841897,2.136667002481654
+320,8,160,80,12800,0.125,5,0.8427154872988737,0.01106454588886201,0.4472135954999579,1.8843691152921422
+320,8,160,160,25600,0.125,5,0.624074713778184,0.005081601989104376,0.31622776601683794,1.9734975256567266
+320,8,320,80,25600,0.125,5,0.6765586923133772,0.007566062117315585,0.31622776601683794,2.1394664384953246
+320,8,320,160,51200,0.125,5,0.47975902055815955,0.003150464325871738,0.22360679774997896,2.1455475655735277